Negative evidence in language acquisition*
Transcript of Negative evidence in language acquisition*
Cognition, 46 (1993) 53-85
Negative evidence in language acquisition*
Gary F. Marcus Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
Received February 7, 1992, final version accepted July 2, 1992
Abstract
Marcus, G.F., 1993. Negative evidence in language acquisition. Cognition, 46: 53-85
Whether children require “negative evidence” (i.e., information about which strings of words are not grammatical sentences) to eliminate their ungrammatical utterances is a central question in language acquisition because, lacking negative evidence, a child would require internal mechanisms to unlearn grammatical errors. Several recent studies argue that parents provide noisy feedback, that is, certain discourse patterns that differ in frequency depending on the grammaticality of children’s utterances. However, no one has explicitly discussed how children could use noisy feedback, and I show that noisy feedback is unlikely to be necessary for language learning because (a) if noisy feedback exists it is too weak: a child would have to repeat a given sentence verbatim at least 8.5 times to decide with reasonable certainty that it is ungrammatical; (b) no kind of noisy feedback is provided to all children at all ages for all types of errors; and (c) noisy feedback may be an artifact of defining parental reply categories relative to the child’s utterance. For example, because nearly all parental speech is grammatical, exact repetitions (verbatim repetitions of child utterances) necessarily follow more of children’s grammatical utterances than their ungrammatical utterances. There is no evidence that noisy feedback is required for language learning, or even that noisy feedback exists. Thus
Correspondence too: Gary F. Marcus, Department of Brain and Cognitive Sciences, Massachusetts
Institute of Technology, Cambridge, MA 02139, USA; e-mail [email protected].
*Many of the ideas in this paper grew out of discussions with Steven Pinker: without him this paper
would not exist. I wish to thank both Steve Pinker and John Rosen for extremely helpful comments at
every stage of the preparation of this manuscript, two anonymous reviewers, S. Avrutin, P. Bloom, S.
Crain, K. Olguin, N. Pearlmutter, D. Poeppel, D. Slobin, E. Stein, and especially Fei Xu, for comments on drafts of this paper, J. Stern and T.J. Rosen for statistical help, and M. Bowerman, E. Clark, H. Clark, S. Prasada and V. Valian for helpful discussion. The author was supported by an
NDSE Graduate Fellowship and NIH Grant HD 18381 and NSF Grant BNS 91-09766 to Steven Pinker.
OOlO-0277/93/$05.00 @ 1993 - Elsevier Science Publishers B.V. All rights reserved.
54 G. F. Marcus
internal mechanisms are necessary to account for the unlearning of ungrammatical utterances.
Introduction
A major challenge for theories of language acquisition is to explain how children
recover from grammatical errors, such as (1) and (2):
(1) ‘1 maked it with water. (Sarah 4;5; Brown, 1973)
(2) *And fill the little sugars up in the bowl. (Mark 4;7; Pinker, 1989)
Do children eliminate their grammatical errors solely on the basis of internal
mechanisms, or do they require external feedback from their parents?
Brown and Hanlon (1970) carried out two analyses in the first systematic
examination of whether parents provide feedback contingent on children’s gram-
matical errors. In the first, they examined whether parents comprehend their
children better if their children’s utterances are grammatical. Each child question
was classified as grammatical or ungrammatical; each parental reply was coded to
indicate whether the parent understood or failed to understand the child’s
utterance. Replies indicating the lack of comprehension followed about as many
grammatical child questions (42%) as ungrammatical ones (47%); similarly,
replies indicating comprehension were equally likely following grammatical and
ungrammatical speech (45% in both cases). Brown and Hanlon concluded that
“In general, the results provide no support for the notion that there is a
communication pressure favoring mature constructions” (p. 45).
In their second analysis, Brown and Hanlon coded whether parental replies
indicated approval (e.g., That’s right or Yes) or disapproval (e.g., That’s wrong or
No). Again, there was no relation between parental reply types and child
grammaticality. Finding not “even a shred of evidence that approval and dis-
approval are contingent on syntactic correctness” (p. 47), Brown and Hanlon
concluded that:
While there are several bases for approval and disapproval, they are almost always semantic or
phonological. Explicit approval or disapproval of either syntax or morphology is extremely rare in
our records and so seems not to be the force propelling the child from immature to mature forms.
(P. 48)
Based largely on their conclusions, much subsequent research in language acquisi-
tion has tried to solve the puzzle of how children acquire language solely from
positive evidence (i.e., hearing sentences that belong to a language) and without
negative evidence (i.e., information about which sentences do not belong to that
language) (e.g., Baker, 1979; Baker & McCarthy, 1981; Berwick, 1985;
Negative evidence 55
Bowerman, 1983, 1987, 1988; Braine, 1971; Chomsky & Lasnik, 1977; Fodor &
Cram, 1987; Grimshaw, 1981; Gropen, Pinker, Hollander, & Goldberg, 1991;
Lasnik, 1981, 1989; Marcus et al., 1992; Matthews & Demopoulous, 1989;
Mazurkewich & White, 1984; Pinker 1984, 1989; Roeper, 1981; Wexler Br
Culicover, 1980; Wexler & Hamburger, 1973; Wexler & Manzini, 1987.)
The “no negative evidence problem” - that is. how children could learn
language without negative evidence - is often seen as necessarily tied to nativist
explanations that posit the existence of internal and innate mechanisms. In fact,
Braine (1971), the first to discuss the implications of the “no negative evidence”
problem, used the problem to argue against Chomsky’s (1965) nativist explana-
tion of language acquisition. Arguing that Chomsky’s hypothesis-testing proposal
would require negative evidence, Braine presented anecdotal evidence suggesting
that negative evidence was insufficiently frequent and perhaps ignored, hence
refuting Chomsky’s proposal.
Without negative evidence, any model of language acquisition, nativist or not,
must account for how the child can learn language from positive evidence alone.
Suppose a child must learn that a certain sentence, B, is ungrammatical. The child
might learn this in two ways. First, something external to the child (i.e., explicit
negative evidence) could tell the child that B is ungrammatical. If the parent does
not provide an explicit denial of sentence B, the only alternative is that the parent
says a sentence (or a set of sentences), A, and that the child has a mechanism
which eliminates B when given A. That mechanism, innate or learned, must be
internal to the child, regardless of whether it is, for example, a general pragmatic
mechanism or a linguistically specific mechanism. The existence or non-existence
of negative evidence allows us to determine whether internal mechanisms are
needed. If negative evidence does not exist, the task of language acquisition
researchers must be to discover which internal mechanisms do allow children to
eliminate their errors.’
Recently, in what I will call discourse studies, some researchers have disputed
the claim that there is no negative evidence (Bohannon & Stanowicz, 1988;
Demetras, Post, & Snow, 1986; Hirsh-Pasek, Treiman, & Schneiderman, 1984;
Morgan & Travis, 1989). These studies, reporting evidence from the distribution
of certain patterns of discourse between parents and children, argue that parents
“‘Indirect negative evidence” (e.g., Chomsky 1981) - information about which sentences have not
appeared in the input - is sometimes offered as a third type of evidence available to the child. The use
of indirect negative evidence would require an inference such as “If X appears in the input, assume
that Y is not in the language, unless you hear Y in the input.” Indirect negative evidence thus depends on a reanalysis of positive evidence based on mechanisms internal to the child, rather than input
external to the child; hence for the purposes of this paper I collapse it with positive evidence. See
Pinker (1989, pp. 14-15) for further discussion.
Marcus 56 G.F
Table 1. Definitions of types of parental replies used in discourse studies
Explicit approval
Non sequiturs
Repetitions (only Hirsh-Pasek et al.)
Imitations/exact reps.
Expansion
Recasts
Topic extensions
Move-onslno responses
Clarification questions
Confirmation questions
Parent says yes or uh-huh or the like
Parent fails to understand child
Parent repeats child utterance, perhaps with changes
Parent repeats child utterance,
verbatim
Parent repeats child utterance, but
makes grammatical changes or adds
new material
Parent repeats child utterances. but
with minor grammatical changes
Continued topic but not a repetition or
expansion
Parent moves conversation along
Parent asks a question that requests
child to repeat part of utterance
Parent asks a yes/no question
do provide implicit feedback to their children based on whether children speak
grammatically.’
Each study used somewhat different reply categories; the Appendix reprints
exact definitions. Table 1 provides rough definitions of each type of parental
reply. Some discourse patterns, such as expansions (interchanges between a child
and a parent in which a parent repeats the child’s utterance with small changes),
are claimed to be elicited more often by ungrammatical speech than by grammati-
cal speech. Other discourse patterns such as exact repetitions (interchanges in
which a parent exactly repeats a child utterance) are claimed to be elicited more
often by grammatical speech than by ungrammatical speech.
In all cases, the differences in parental replies to grammatical versus un-
grammatical sentences were statistical rather than categorical (i.e., all-or-none). I
call this type of feedback noisy feedback, since parents provided each type of
‘All of the studies that I will discuss contrast parental replies to grammatical sentences with
parental replies to ungrammatical sentences. However, one study, Moerk (1991), presented examples
only of how parents reply to ungrammatical sentences and failed to present examples of how parents
reply to grammarical utterances. For example, in the following exchange, Moerk called the reply What
did I do? a “correction” (Eve: What did you doed? Mother: What did I do?), but he neglected to note that the identical parental reply, two samples earlier, is not a “correction” (Eve: Wurt my eye.
Mother: Your eye? Eve: Yeah. Mother: What did I do?) (Eve, 2;l). Moerk provided no objective definition of “corrections” and no explanation of how a child could recognize or use them. Applying
this method, one could argue that the word red provides negative evidence: simply cull transcript
examples in which the parent said red following a child’s ungrammatical utterances and ignore the
occasions in which the parent said red after the child’s grammatical utterances. Without the crucial contrast between replies to grammatical and ungrammatical sentences, such studies are irrelevant to
the question of whether parents provide negative evidence.
Negative evidence 57
reply after both grammatical and ungrammatical sentences, albeit in different
proportions.
In a recent discourse study, Bohannon and Stanowicz (1988) argue that
discourse-based noisy feedback “may be considered superior to simple denials
and as qualified to assume the role of ‘negative evidence’ ” (p. 688), and
concluded that “To the extent that current theories of language acquisition also
ignore adults’ tendency to provide feedback (i.e., negative and specific evidence),
these theories will fail to accurately account for language acquisition” (p. 688).
These conclusions, if correct, would undermine much of the research based on
the assumption that parents do not provide negative evidence. However, because
the methodology and results of the discourse studies vary and often conflict with
one another, and because there has been little discussion of how children could
use noisy feedback to eliminate grammatical errors, the role and existence of
noisy feedback remain controversial. This controversy is evident both empirically,
in the conflicting results and conclusions of Bohannon and Stanowicz (1988) and
Morgan and Travis (1989), and theoretically, in a recent debate between Gordon
(1990) and Bohannon, MacWhinney, and Snow (1990).
The goal of this paper is to define and apply clear criteria for clarifying the role
and existence of noisy feedback. These criteria will lead to the following
conclusions:
. Noisy feedback is too weak to be a plausible way of eliminating errors. Even
under statistically optimal conditions, a child would have to repeat a given
sentence verbatim more than 85 times to eliminate it from his or her grammar.
. Noisy feedback is inconsistent across parents, declines or disappears with age,
and is probably not provided for all types of errors and is thus unavailable for
much of language acquisition. Furthermore, some apparent patterns of noisy
feedback may be averaging artifacts that do not correspond to types of feedback
given to any individual child.
. Because parental reply categories are defined only with respect to the child’s
utterance, “correlations” between the two may be artifacts resulting from the
definition of parental reply categories and of constant noncontingent properties
of parental and child speech. For example, several studies show that verbatim
repetitions follow more grammatical than ungrammatical sentences. But be-
cause parents virtually never speak ungrammatically, verbatim repetitions
necessarily negatively correlate with children’s ungrammatical utterances.
Any one of these three conclusions, the weakness, the inconsistency, or the
inherently artifactual nature of noisy feedback, would by itself severely under-
mine, if not falsify, the position that parents provide feedback to their children
and thus that such feedback is required for language acquisition. Of course, even
if parents provided feedback, it might turn out empirically that children do not
use it. For example, Zwicky (1970), in discussing his daughter’s inflectional
58 G. F. Marcus
overregularizations (e.g., goed), points out that “six months of frequent correc-
tions by her parents had no noticeable effect”. Braine (1971) even more clearly
illustrates that feedback may be ineffective:
For experimental purposes, I have occasionally made an extensive effort to change the syntax of
my two children through correction. One case was use by my two-and-a-half-year-old daughter of
other one as a noun modifier. Over a period of a few weeks I repeatedly but fruitlessly tried to
persuade her to substitute other + N for orher one + N. With different nouns on different
occasions, the interchanges went somewhat as follows: “Want other one spoon, Daddy” - “You mean, you want THE OTHER SPOON” - “Yes, I want other one spoon, please, Daddy” - ‘Can
You say ‘the other spoon’?” - “Other one spoon” - “Say ‘other’ ” - “Other” -
“Spoon” - “Spoon” - “Other spoon” - “Other. spoon. Now give me other one spoon?” Further tuition is ruled out by her protest, vigorously supported by my wife. Examples indicating a
similar difficulty in using negative information will probably be available to any reader who has
tried to correct the grammar of a two- or three-year-old child. (pp. 160-161)
Even when children understand that they are being corrected, they sometimes
make incorrect generalizations, as McNeil1 (1966, p. 69) shows:
Child: Nobody don’t like me.
Mother: No. say “nobody likes me.”
Child: Nobody don’t like me.
[Eight repetitions of this dialogue follow.]
Mother: No, now listen carefully, say “NOBODY LIKES ME.”
Child: Oh! Nobody don’t likes me.
These examples suggest that it is the child’s underlying linguistic system, rather
than negative evidence, which forces children to change their grammars. Until the
existence of reliably parental feedback is firmly established, however, it is
premature to consider whether children actually use it.
Types of parental feedback and their role in language learnability
Positive evidence is simply the input, that is, the sentences children hear. In
contrast, negative evidence is a parental behavior that provides information about
when sentences are not in the language. Negative evidence does not tell a child
which sentences are grammatical; rather, it indicates that the child has uttered an
ungrammatical sentence. Moreover, negative evidence does not tell a child why a
particular sentence is ungrammatical.
A child who does not speak can receive no negative evidence, aside from
explicit metalinguistic statements (e.g., don’t say X). Negative evidence must
respond to a child’s utterance; it can only arise if parents are in some way sensitive
(though not necessarily consciously) to whether their children speak grammati-
tally The parental behavior that provides negative evidence I will call the reply type.
The reply type may take many forms: the parent might say 110, provide a
Negative evidence 59
repetition, shrug, or even spank the child. An individual occurrence of a given
reply type (e.g., a single utterance of the word no) is a repfy instance. To distinguish three kinds of negative evidence that are often collapsed, I use
the terms complete feedback, partial feedback and noisy feedback. Complete feedback is any corrective signal provided for all and only those sentences that are
ungrammatical (cf. Gold’s (1967) term informant presentation in the formal
mathematical literature on learnability). Partial feedback is any corrective signal
provided following some ungrammatical sentences, but never provided following
correct sentences. Noisy feedback is a corrective signal provided after some errors
and after some correct sentences, but in different proportions. A single reply
instance of noisy feedback, unlike instances of complete feedback and partial
feedback, does not guarantee that a sentence is ungrammatical.
Verbal parental feedback, such as repetitions or expansions, inherently pro-
vides the child with positive evidence, as well as potentially providing negative
evidence. Several studies have reported that expansion may spur learning (though
Morgan, Bonamo, & Travis, 1991, conclude the opposite). Crucially, however,
even if children learn from expansions, they might do so without using negative
evidence. In principle, a child could learn whether some sentence is grammatical
(a) only from positive evidence, which tells the child that the parental utterance is
a possible way of saying what she wanted to say; (b) only from negative evidence,
which tells the child that her utterance is not acceptable; or (c) both, which would
tell the child both that her sentence was unacceptable and provide an acceptable
alternative; each of these possibilities is plausible a priori. Consider this hypo-
thetical dialogue:
(3) Child: I eated the food.
Parent: I ate the food.
This parental reply clearly provides the child with a piece of positive evidence: I
ate the food is a grammatical sentence. Positive evidence alone does not tell the
child whether eated and ate are stylistic variants or synonyms or whether eated is
unacceptable, but might do so in combination with internal mechanisms. The
parental reply could serve as negative evidence only if the child recognizes it as an
exemplar of a recast and has mechanisms that use such information.
Complete feedback
A child who received complete feedback might eventually converge on the correct
language by eliminating all possible but ungrammatical sentences from her
language (Gold, 1967; Pinker 1979).
If complete feedback existed it would probably be easy to detect, since there
would, by definition, be a perfect correlation between the reply type and the
60 G. F. Marcus
grammaticality of the child’s utterance. Nobody, though, has ever provided any
evidence that children receive complete feedback; every reply type studied in the
discourse studies is not provided for all ungrammatical sentences, and is provided
for some grammatical sentences. Nobody claims, for instance, that parents
expand children’s utterances every time that a child errs but never when the child
speaks grammatically. Thus children almost certainly do not receive complete
feedback.
Partial feedback
Partial feedback, by definition, is not provided every time a child makes an error.
There has been no explicit discussion of how children could use such feedback,
although Gordon (1990) has suggested that “it may, in fact, be possible to show
that learning can occur with inconsistent feedback . . . using statistical rather than
absolute criteria to determine grammaticality” (p. 219). It is not clear whether
children have mechanisms for using partial feedback, nor what those mechanisms
might be.
In any case, partial feedback does not seem to exist. No reply type studied in
the discourse studies is provided only for ungrammatical sentences but not for
grammatical sentences, or vice versa. The evidence from the discourse studies
suggests that all patterns of discourse provide, at best, noisy feedback: every reply
type is provided following both grammatical and ungrammatical speech.
Noisy feedback
Several studies claim to have documented patterns of noisy feedback, but there
has been no explicit discussion of how children might use it. I will examine a
concrete example, after defining four terms. The proportion of reply instances
(e.g. expansions) elicited by grammatical speech will be denoted by pr{gr}. The
proportion of reply instances elicited by ungrammatical speech is, then, pr{ungr}.
The number of times the child tests a given sentence is n; the observed proportion
of reply instances is p(obs). I will assume that the feedback is provided prob-
abilistically and independently.
Suppose pr{ungr} = .20 and pr{gr} = .12 (These proportions are drawn from
the distribution of noisy feedback reported by Hirsh-Pasek et al.) Suppose the
child says a sentence once (i.e., II = 1). Depending on whether the parent
provides a reply instance, p(obs) will be either 0 or 1. Suppose p(obs) is 1; given
the small sample size, p(obs) could easily come from either distribution. The
Negative evidence 61
chance that p(obs) comes from pr{ungr} is .625 (since 20 out of every 32 reply
instances would follow ungrammatical sentences). The chance that it comes from
pr{gr} is ,375 (=l - .625). Thus if the child decides that the sentence is
ungrammatical, there is a .375 chance that the child is wrong. Or if p(obs) is 0,
there is a .52 (881168) chance that the sentence is grammatical, and a .48 chance
that the sentence is ungrammatical. In either case, the child has some information
but not enough.
The child’s chances improve if the child increases the sample size, n, by
repeating the sentence. Suppose after 100 repetitions of the sentence, p(obs) =
.17. Then it is relatively likely that the sentence is from pr{ungr} but still
reasonably likely that p(obs) comes from pr{gr}. If a child repeats the test
sentence 1000 times, with p(obs) = .20, the child could decide with near certainty.
In general, the more times the child repeats the same sentence, the more
information the child gathers, and the easier it is for the child to determine the
grammaticality of that sentence, with some degree of certainty; I will assume that
the degree of certainty equals the rate of error.
Because the end point of language acquisition is the adult state, I assume that
the acceptable rate of error, F, should reflect the rate at which adults make
grammatical errors (or perhaps be lower, since adults often know when they have
made errors). If E is too high, the child will fail to converge on the adult state of
rarely making errors. The overall rate of adult past tense overregularization errors
(e.g. maked or maded) is around .00004 (Marcus et al., 1992), while the overall
rate of adult speech errors is about .OOl (Stemberger, 1989). To be generous by
an order of magnitude, suppose E is .01.3
In Figure 1, the left curve represents the sampling distribution of the propor-
tion of parental reply instances out of child utterances provided following
grammatical sentences, while the right curve represents the sampling distribution
of the proportion of parental reply instances out of child utterances provided
following ungrammatical sentences. The shapes of the two curves depend on the
proportions of feedback to grammatical and ungrammatical sentences and the
number of times, n, that the child repeats the test sentence; here pr{gr} = .12,
pr{ungr} = .20, y1= 10.
In order to decide whether a sentence is grammatical, assuming E = .Ol, a child
must determine whether p(obs) is within the 99% confidence interval of the
grammatical distribution, within the 99% confidence interval of the ungrammati-
cal distribution, or within the overlap between the two confidence intervals, which
I call the region of ambiguity, shown as the cross-hatched region in Figure 1. This
‘Actually, two error rates are involved here: the chance of incorrectly accepting a sentence as grammatical and the chance of incorrectly accepting a sentence as ungrammatical. Since most speakers
agree both on which sentences are grammatical and on which sentences are ungrammatical, both rates should be very low. Lacking evidence to the contrary, I assume that these two error rates are equal.
62 G. F. Marcus
Figure 1.
PI+}= -12 Prbngr}=.20
OfI 0.1 02 03 OA 05 0.4 0.7 0.8 0.9 la
p(obs) Obsewed proportion of reply instances out of child utterances of a given sentence
Proportion of child utterances of a given sentence which are followed by a reply instance, given n = IO. The vertical line, at p(obs) = .36, represents the upper bound of the 99% confidence interval of pr{gr). The cross-hatched region of ambiguity represents the range of possible proportions of feedback which fall within the 99% confidence intervals of the distributions of parental responses both to grammatical and ungrammatical sentences. The dotted region to the right represents the area of confidence interval of pr{ungr} that is not within the confidence interval of pr{gr}.
region indicates possible values of p(obs) under which a child cannot make a
decision within the specified risk of error.4
The only way the child can reduce the region of ambiguity is to repeat her test
sentence again and again, n times. Figure 2 represents the smallest value of n in
which the region of ambiguity is eliminated (for F = .Ol). Before being able to
make a decision, the child would need to repeat the same construction 446 times.
Of course, the minimum number of repetitions necessary to decide whether a
sentence is grammatical depends on the proportions of feedback; these values
‘Note that for positive evidence there is virtually no region of ambiguity. If the parent says a
sentence, the child can be nearly certain that the sentence is grammatical. According to Newport.
Gleitman, and Gleitman (1977), 99.44% of parental speech to children is grammatical. Similarly, if the parent provided complete feedback, there would also be no region of ambiguity. Upon hearing a
reply instance, the child could immediately decide that a given sentence was ungrammatical.
Negative evidence 63
PI-&r}=_12 Prhmgr}=.20
1
03 OR 05 Ob 0.7 08 0.9 lfl
pbbs)
Obsewed proportion of reply instances out of child utterances of a given sentence
Figure 2. Proportion of child utterances of a given sentence which are followed by a reply instance, given n = 446. The vertical line, at p(obs) = .16, represents both the lower bound of the 99% confidence interval of pr{ungr} and the upper bound of the 99% confidence interval of pr{gr}. Thus, the region of ambiguity is zero. The left region indicates the confidence interval of pr{gr}. The right, dotted region, represents the confidence interval of pr{ungr}.
vary across the discourse studies. Table 2 shows estimates of y1 for the clearest and
easiest reply types the child could use. Weaker differences between parental
replies to grammatical and ungrammatical sentences would require the child to
repeat a test sentence even more often.
The lowest estimate of II, derived from data reported by Bohannon and
Stanowicz (1988), is 85; the highest estimate of II is 679. If a child does not repeat
a given sentence n times, noisy feedback cannot account for how the child
eliminates the error.
However, aside from formulas and routines, most of children’s speech is not
repetitive. Children repeat few sentences even ten times, let alone 85. For
example, Pinker (1989) searched 86,332 child utterances taken from transcripts of
spontaneous speech for errors in which a child misused the argument structure of
a dative (e.g., Don’t say me that). None of these errors was repeated anywhere
near 85 times: one child (Eve) made 11 errors with the verb write, no other error
64 G.F.
Table 2. Minimum number of times (n) a child would need to repeat a given
sentence verbatim to decide whether it was grammatical, with chance of
error, E, less than .Ol, calculated for the strongest reply types reported in
the discourse studies (n may be determined by finding the minimum
number of repetitions of a single sentence such that the 99% confidence
intervals of the two distributions (pr{gr}, pr{ungr}) do not overlap)
Study
Hirsh-Pasek et al.
Penner
Bohannon and Stanowicz
Morgan and Travis
Reply type prfungrl pr{gr) n
Repetitions 2-year-olds 20 12 446
3-S-year-olds no differences n/a Expansions: 1 group 18.3 4.6 104
Expansions: group 2 11.3 6.3 679
Total of all reply types 35 14 85
Expansions (A. E. and S) 11.3 2.7 169
Note: The data from Demetras et al. are excluded because most patterns of reply varied
across parents and no inferential statistics (or sample sizes) were provided. Without such
information we cannot infer the reliability of the data. The data for Morgan and Travis
are an average across the patterns of parental expansions to Adam, Eve. and Sarah,
from Table 4. Morgan and Travis (1989, p, 545). because there is substantial variation
between the pattern of expansions to the different children. Eve would only require 63
verbatim repetitions, but Sarah would require 300.
was repeated more than three times, and thus noisy feedback is almost certainly
too weak to account for how children eliminate their errors.
Children might follow strategies other than the one I have outlined, but none
of these can reduce n without increasing the error rate. For instance, children
might be thoroughly productive and assume that every sentence is grammatical
unless the sentence receives frequent and repeated correction. Following this
strategy, children would overgenerate implausibly often, repeating ungrammatical
sentences far more often than they are actually observed to do (Marcus et al.,
1992). Children also cannot be completely conservative and assume that all
sentences are ungrammatical, since they productively create sentences they have
never heard (e.g., Pinker, 1989). In any case, if children utilize noisy feedback
before repeating a sentence enough times, their risk of errors must go up, because
there is only a limited amount of information in each reply instance, and the only
way to get more information is to repeat their sentences again and again in an
effort to increase the reliability of the information contained in parental replies.
Combining noisy feedback from different sentences
I have assumed that children determine grammaticality on a sentence-by-
sentence basis. Could children instead determine the grammaticality of entire
classes of sentences on the basis of negative evidence? For instance, a child trying
to test whether The boy eat apples is grammatical might combine feedback
Negative evidence 65
provided to that sentence with feedback provided to “equivalent” sentences such
as The girl eat apples, in order to overcome the inherent weakness of noisy
feedback.
Combining feedback provided to some set of equivalent sentences requires a
child to determine which sentences are “equivalent”; determining equivalency,
however, can only be done properly after the child knows which factors affect a
sentence’s grammaticality, defeating the very purpose of using noisy feedback.
Sometimes two sentences similar in structure, differing by no more than a word,
are both equally acceptable in a language; for example, A boy ate some apples and A girl ate some apples are both grammatical, and Some a boy apples ate and
Some a girl apples ate are both ungrammatical. Often, however, two apparently
quite similar sentences are not equally acceptable in the target language. For
example, A boy ate some apples is grammatical but A boy ate much apples is
ungrammatical. A child who combined the feedback provided for these two
sentences would err, as would a child who combined feedback for *Fill the water in the bowl with Pour the water in the bowl, or John slapped Bill with *John hitted Bill. The rules governing which sentences are and are not equivalent, with respect
to feedback, are the very rules of the grammar the child is trying to learn. If
feedback is to be the means by which children form equivalence classes then
children’s use of feedback cannot rely on the pre-existence of those equivalence
classes. There is no doubt that children do form equivalence classes, but negative
evidence does not explain how children form them.5
Learning the probabilities of noisy feedback
The model presented above assumes that the child has perfect knowledge of
(pr{gr} and pr{ungr}). If the child’s knowledge of the reply probabilities is poor,
errors of falsely accepting ungrammatical sentences or else errors of falsely
rejecting grammatical sentences becomes more likely. For instance, if pr{gr} =
.12 and pr{ungr} = .2, but the child incorrectly hypothesizes that pr{gr} = .2 and
pr{ungr} = .4, and the child utters a grammatical sentence multiple times and
receives p(obs) = .2, the child will conclude falsely that the sentence is gram-
matical.
Given individual differences between parents (to be discussed below), it is
unlikely that the probabilities of feedback are specified innately. How, then,
could the child learn the reply probabilities? The child cannot simply record how
many of her ungrammatical versus grammatical sentences elicit instances of some
reply. Children could safely assume that their parents speak grammatically, but by
‘The child learning from positive evidence alone may seem to face the same dilemma- how to form the equivalence classes without knowing them. The answer proposed by Chomsky and others is
that the child has innate knowledge which constrains the inductions by which the child may form
equivalence classes.
hypothesis they do not know which sentences are ungrammatical. A child might
construct minimal pairs by comparing the parent’s sentence with a sentence that is
likely to be ungrammatical. But this begs the question: how would the child know
which sentences are likely to be ungrammatical? In sum, it is unclear how children
could learn the probabilities of feedback.
Some researchers seem to want to eliminate language-specific learning mecha-
nisms in favor of parental feedback on the grounds that models that depend on
parental feedback might be simpler, but consider what new mechanisms they must
propose. The child must recognize the (subtle) feedback reply types and discrimi-
nate the useful reply types from the noise, and then make use of them.
Furthermore, the child must learn (at least approximations of) the probabilities of
feedback, induced in some unspecified manner. Though a child might solve all
these problems, it is hardly clear that models of language acquisition that depend
on feedback present simpler models than those based on internal, linguistic-
specific constraints.
Summary
I distinguished three types of negative evidence: complete, partial, and noisy
feedback. There is no evidence that either complete feedback (negative evidence
provided following all ungrammatical sentences and no grammatical sentences) or
partial feedback (negative evidence provided for some ungrammatical sentences
but no grammatical sentences) exists. Noisy feedback, the only attested type of
feedback, is negative evidence provided for some grammatical sentences as well
as some ungrammatical sentences. Under conditions of noisy feedback (unlike
complete feedback and partial feedback), a single reply instance cannot guarantee
that a given sentence is ungrammatical. Under certain reasonable assumptions
about feedback probabilities and acceptable error levels, a child would have to
repeat the sentence in question, verbatim, at least 85 times before deciding
whether it is grammatical. Furthermore, to use noisy feedback, the child must
solve the difficult bootstrapping problem of determining the probabilities of reply
instances for both grammatical and ungrammatical sentences. Theories of lan-
guage acquisition that require noisy feedback are thus plausible only to the extent
that it is plausible that children notice noisy feedback, have mechanisms to make
use of it, and repeat sentences at least 85 times each.
Generality and availability of noisy feedback
Noisy feedback is interesting only if it can eliminate the need for specific internal
mechanisms, but these mechanisms are only eliminable if noisy feedback is
Negative evidence 67
necessary for language learning. If noisy feedback were merely helpful for
language learning, but not necessary, internal mechanisms would still be needed
to account for language acquisition. In this section I will develop and apply
criteria, based in part on those discussed by Grimshaw and Pinker (1989), Pinker
(1989), and Morgan and Travis (1989), for determining whether there is evidence
that any type of feedback is a necessary component of language acquisition.
Criteria
Does every child receive noisy feedback?
Since every child learns language, a particular type of parental feedback can be
a precondition for learning language only if it is provided to every child. Many
types of feedback may be sufficient for language acquisition, but if any child
manages to learn language lacking that particular type of feedback, then that
particular type of parental feedback cannot be necessary for language acquisition.
(Note that every child receives positive evidence.)
Because between-subject differences are crucial, appropriate statistical analy-
ses of patterns of noisy feedback should examine patterns of feedback within
individual parent-child dyads. Combining data from several children may create
two types of averaging artifacts. First, aggregating data from multiple children
may hide differences between the patterns of feedback provided to those children.
Suppose one parent provided reliable noisy feedback, but another parent did not
provide reliable feedback. Combining the data from these two parent-child
dyads, one might falsely conclude that the pattern of feedback is available to both
children. To avoid such problems, a more fine-grained analysis is necessary.
Second, when studies average raw data6 from several dyads, patterns of noisy
feedback can emerge that no child actually received. For example, if 80% of
Johnny’s utterances are grammatical, and Johnny’s mother repeats a random 75%
of his utterances, but 60% of Billy’s utterances are grammatical and his mother
only repeats half of Billy’s sentences, at random, then combining the same
amount of data for both children will lead to the (false) conclusion that 64% of
grammatical utterances elicited repetitions, compared to only 58% of ungram-
matical utterances; see Table 3.
Do parents provide noisy feedback until errors stop?
If certain aspects of language acquisition depend on parental feedback, parents
must provide feedback until children have finished acquiring those aspects of
bAveraging percentages rather than raw data would prevent the emergence of entirely spurious
reply contingencies but could still conceal important individual differences.
68 G. F. Marcus
Table 3. How patterns of feedback could be an artifact of averaging data from several children
Johnny
Billy
Average
Grammatical
60180 = .75
30160 = .5
901140 = .64
Ungrammatical
15120 = .75
20/40 = .5
35160 = .58
Now: Number of utterances repeated out of 100.
language. Moreover, if feedback declines or disappears with age, combining data
might suggest falsely that noisy feedback is available, even if no child actually
receives such feedback. Suppose that a parent consistently repeats every utterance
of his or her child at age 2 but by the time the child is 3 the parent no longer
repeats many of the child’s utterances. If at age 2 the child produces 80%
ungrammatical utterances, but by age 3 the child only produces 60% ungrammati-
cal sentences, then, as shown in Table 4, one might conclude falsely that the
parent repeats a higher proportion of ungrammatical sentences than grammatical
sentences, even though no parent ever replies differentially to grammatical and ungrammatical sentences. The misleading conclusion results from (a) the parent’s
repeating fewer of the older child’s sentences and (b) the older child’s greater
proportion of grammatical sentences. Averaging data across different children of
different ages may cause similar artifacts.
Do parents provide noisy feedback for all kinds of linguistic errors?
Language has several components or levels of representation, including
phonological, syntactic, and semantic representations. There is empirical evidence
that parents reply to different types of errors in different ways. For example,
Brown and Hanlon (1970, p. 48) concluded that the “bases for approval and
disapproval. . are almost always semantic or phonological”. To show that
negative feedback is necessary for the elimination of errors of some kind,
phonological, morphological, syntactic, or semantic, one must provide evidence
that adequate feedback corresponding to that specific kind of error exists.
Table 4. How patterns of feedback could be an artifact of averaging over age
Grammatical Ungrammatical
2-year-old 20/20 = 1 80180 = 1
3-year-old 0140 = 0 0160 = .O
Average 20160 = .33 801140 = .53
Note: Number of utterances repeated out of 100.
Negative evidence 69
Do parents provide noisy feedback for different types of grammatical constructions?
Parents might provide feedback for some syntactic constructions, but still fail
to provide feedback for other syntactic constructions. A child who received
parental feedback for, say, agreement errors (e.g., He eat the candy), but not for
improper locative verb alternations (e.g., Fill the little sugars up in the bowl), could not learn which locative alternations are permissible on the basis of parental
feedback. Parental feedback must be available for all grammatical constructions
for which it is posited to correct.
Summary
To determine whether a particular type of noisy feedback could be necessary for
unlearning errors, there are four criteria: (1) the reply type must be available to
all children; (2) the reply type must be available throughout acquisition; (3) the
reply type must be available for errors in each component of language; (4) the
reply type must be available after all types of errors within a given component.
Any pattern of noisy feedback that is necessary for language acquisition would
meet all of these criteria. Positive feedback meets all of them: all children, at all
ages, receive positive evidence for all types of representation and nearly all types
of constructions. Complete feedback, if it existed, would also meet these criteria.
Application of criteria
Does every child receive noisy feedback?
Many patterns of feedback were inconsistent across dyads within the studies
that reported them, as indicated by asterisks in Table 5. Futhermore, many of the
remaining effects failed to replicate in other studies.’ The rightmost column
indicates that only one pattern of feedback was provided to all children studied:
recasts. Bohannon and Stanowicz found that, on average, parents provided more
recasts after ungrammatical than grammatical sentences. However, though they
indicate that parents and non-parents replied differently, it is unclear whether
every child received more recasts after ungrammatical sentences; even recasts
may not be consistent across the children they studied.
The studies that allow examination of individual differences show that different
parents provide different, even conflicting, patterns of feedback. Demetras et al.
‘An anonymous reviewer suggested that differences in social or economic status might account for some of the variation in feedback.
70 G. F. Marcus
Table 5. Availability of particular types of feedback across all children studied
Did all children
receive the same
type of Found by Not found by feedback’?
Explicit approval BH. HPTS, P No Non sequiturs BH No
Repetitions HPTS DPS* No
Imitations (or exact repetitions) DPS, BS P, MT* No
Expansion P, BS DPS*, MT No
Recasts BS Yes
Topic extensions/move-ons DPS P,* MT* No Clarification questions DPS, BS MT No
Confirmation questions P, MT No
Notes: Morgan and Travis distinguished partial and exact imitations. Exact imitations
were consistent across parents; partial imitations were not.
Penner found that confirmation questions were more common after ungrammatical
sentences; these effects, however, were not statistically significant.
*Indicates that the effect found was inconsistent across children within that study.
BH, Brown and Hanlon; HPTS, Hirsh-Pasek et al.; DPS, Demetras et al.; P, Penner;
BS, Bohannon and Stanowicz; MT, Morgan and Travis.
found that some children receive certain reply types, such as expanded repeti-
tions, more after ill-formed sentences than well-formed sentences.’ However,
other children receive the same reply types more often after well-formed sen-
tences. Penner noted that “although parents repeated correct child utterances
slightly more frequently than incorrect child utterances, the pattern was not
consistent for all parents”. Morgan and Travis, studying three subjects, Adam,
Eve, and Sarah (from Brown, 1973), found that, for each set of parents, the
contingencies between parental replies and the grammaticality of children’s
utterances were different. For example, while Adam’s parents asked more
clarification questions after his well-formed wh- questions, Eve and Sarah’s
parents were more likely to ask clarification questions after their children’s
ill-formed wh- questions. If the reply types are contingent on grammaticality for
some children, but contingent on ungrammaticality for other children, would
some children eliminate goed in favor of went, while other children eliminate went
in favor of goed?
Finally, Hirsh-Pasek et al., Penner, and Bohannon and Stanowicz all combined
data from many children in their analyses,” leading to potential statistical artifacts.
Data from children who received no feedback, when averaged with data from
‘Demetras et al. presented no inferential statistics, and many patterns of feedback are different for different parent-child dyads. Thus contingencies within individual dyads may be due to chance or
sampling error. “Penner provided information about the reliability of patterns of feedback across parents, but the
other two studies do not.
Negative evidence 71
children who did receive feedback, might incorrectly suggest that feedback is
available to all children. Moreover, as shown above in Table 4, it is even possible
that some alleged patterns of feedback were not provided to any children.
In fact, there may be no type of feedback that is provided universally. Gordon
(1990) argued that in the Piedmont Carolinas, and possibly in other cultures,
parents rarely speak directly to their children. The existence of cultures in which
parents do not provide any feedback would prove that children can learn from
positive evidence alone and that no form of feedback can be a necessary
precondition to learning language.
Bohannon et al. (1990, p. 224) criticize Gordon’s example as being “an
anecdote . . . [that] hardly constitutes believable counterevidence”. In fact, Gor-
don’s example is based on a well-documented ethnographic study from Heath
(1983), from which I quote:
Trackton adults do not see babies or young children as suitable partners for regular
conversation [Ulnless they wish to issue a warning, give a command, provide a recommenda-
tion. or engage the child in a teasing exchange, adults rarely address speech specifically to very
young children. (p. X6)
Many other cultures with markedly different patterns of parent-child interaction
exist. Ochs and Schieffelin (1984) discuss two such cultures. (See also Eisenberg,
1982; Pye, 1986.) In Kaluli, extensive adult modeling is reported, along with a
lack of expansions - apparently without any evident consequences for either rate
or success of acquisition:
In addition to instructing their children by telling them what to say in often extensive interactional
sequences, Kaluli mothers pay attention to the form of their children’s utterances. Kaluli correct
the phonological, morphological, or lexical form of an utterance, or its pragmatic or semantic
meaning. (p. 293)
Rather than offering possible interpretations or guessing at the meaning of what a child is saying, caregivers make extensive use of clarification requests such as “huh?” and “what?” in an attempt
to elicit clearer expression from the child. However, caregivers do not elaborate or expand
utterances initiated by the child. (p. 294)
In the case of Samoan, parental feedback is minimal - again, with no apparent
consequences for acquisition:
[Claregiver speech is largely talk directed a/ the infant and typically caregivers do not engage in
“conversations” wirh infants over several exchanges. When a small child begins to speak, he or
she learns to make his or her needs known to the higher ranking caregiver. The child learns not to
necessarily expect a direct response. (p. 296)
Procedures for clarification are sensitive to the relative rank of conversational participants in the
following manner. If a low status person’s speech is unclear, the burden of clarifications tends to
be placed more on the speaker. [This] situation applies to most situations in which young children produce ambiguous or unclear utterances. Both adult and child caregivers tend not to try to
determine the message content of such utterances by, for example, repeating or expanding such an utterance with a query intonation, A caregiver may choose to initiate clarification by asking
72 G. F. Marcus
“What?” or “Huh?” but it is up to the child to make his or her speech intelligible to thk addressee.
(P. 298)
Even if Bohannon et al. concede such examples, they still argue that
The absence of a particular form of feedback in a particular community does not belie its utility for
those children who do receive it, nor does it mean that no form of feedback is necessary for
language learning to proceed normally. (1990, p. 224)
Although it is true that some form of feedback may be sufficient for those who
receive it, any type of feedback that is not available universally cannot be a
necessary type of feedback. (This does not eliminate the possibility that some
other form might be necessary.)
Could different children use different reply types to learn language? In that
case, each child would need to determine which types of feedback his or her
parents provide. One might argue that children attend to all possible reply types,
and that the reply types that their parents do not use in a manner contingent on
grammaticality have no net effect on the child’s grammar. But if, for example,
some parents use repetitions to indicate grammaticality whereas other parents use
repetitions to indicate ungrammaticality, then each child would have to determine
whether his or her parents’ repetitions correlate positively, negatively, or not all
with grammaticality.
Determining the status of possible reply types raises three problems. First, the
child can only calculate a given reply type’s status if the child knows which
sentences are grammatical. But this is precisely the knowledge the child is trying
to acquire. Second, the bootstrapping problem - determining the probabilities for
a particular reply instance without knowing which sentences are ungrammatical -
is worse if the child does not know which reply types are and are not contingent
on grammaticality. As the child tests more reply types for possible sources of
feedback, the child’s chance of falsely concluding that some reply type is
contingent on grammaticality increases (just as running many t tests increases the
chance of a type I error). Finally, learning which reply types are useful may
require the child to have an unrealistically large memory and unrealistically
powerful processing capacities, because the child would have to record whether
each possible reply instance occurred for every sentence or at least for enough
sentences to discover reliably the reply instance probabilities of a given type of
feedback for a grammatical versus an ungrammatical sentence.
In sum, the only plausible way that some pattern of noisy feedback could be
necessary for language acquisition is if every parent provides the same reply type
to every child, and the child is predisposed to expect that reply type to be
contingent on grammaticality. But the empirical evidence, both within the
discourse studies and across cultures, strongly suggests that no single pattern of
feedback is available to all children, and hence noisy feedback is unlikely to be
necessary for language acquisition.
Negalive evidence 73
Do parents provide noisy feedback until errors stop?
Although several studies appear to show that noisy feedback is available to
2-year-olds, they also show that noisy feedback diminishes rapidly or even
disappears altogether. Hirsh-Pasek et al. found that noisy feedback disappears with age: “. . . . sensitivity to well-formedness is only apparent among [parents of]
2-year-olds. The same pattern of results failed to emerge at other ages” (p. 86).
Penner studied two groups of children, and the older group (mean age = 3;O) received weaker feedback on all measures that the younger group (mean age =
2;O). Morgan and Travis found that parental feedback provided to Eve declined
significantly from the start of data collection (1;lO) to the end of data collection
(2;3). Neither Adam nor Sarah showed significant parental reply contingencies
after age 4. In sum, there is no evidence that children over 4 receive any noisy
feedback.“’
Yet children continue to learn language after feedback has disappeared. Many
types of linguistic errors persist throughout the preschool years. Sarah said I
maked it with water at age 4;5 (Brown, 1973). She makes these inflectional errors
at a greater rate at age 5 than at age 2;6 (Marcus et al., 1992). These errors
continue into at least first grade and perhaps as late as 9 or 10. Similarly, locative
errors, such as Mark’s And fill the little sugars up in the bowl (from Pinker, 1989)
occur until at least age 7 (Bowerman, 1988). Thus noisy feedback, given its
decline over time, is extremely unlikely to account for the unlearning of many of
these errors.
Do parents provide noisy feedback for all kinds of linguistic errors?
Hirsh-Pasek et al., Demetras et al., and Penner collapsed errors of many
linguistic components (e.g., syntax, phonology, and semantics), making it imposs-
ible to distinguish whether parents provide feedback for all types of errors or only
for some types of errors. Hirsh-Pasek et al. and Demetras et al. both collapsed
errors of phonology (e.g., evelator instead of elevator) with syntactic errors in
which words were misplaced or left out entirely. Penner combined morphological
errors (e.g., I ringed the bell) with syntactic errors (e.g., I rang bell), and perhaps
phonological errors as well.” Bohannon and Stanowicz distinguished phonological
errors from syntactic errors but collapsed morphological and syntactic errors.
These studies thus shed little light on whether noisy feedback is provided for
different types of errors (c.f. Valian, in press).
‘“Furthermore, Bohannon and Stanowicz averaged data from different children of different ages
and thus some alleged patterns of feedback might be the result of averaging artifacts. “It is unclear from Penner’s text how phonological errors were treated, but Bohannon and
Stanowicz (1988) write that “Penner (1987) lumped pragmatic, semantic, syntactic, and phonological
errors into a single category of ill-formed children’s speech” (pp. 684-685).
14 G. F. Marcus
The very limited data available suggest that it is unlikely that parents provide
adequate feedback for many types of errors. Bohannon and Stanowicz found that
parents are much more likely to correct semantic errors than syntactic errors
(88.6% vs. 35.9%). Morgan and Travis, the only researchers to separate particu-
lar types of grammatical errors, found that none of the children they studied
received any statistically valid reply types contingent on the grammaticality of
specific types of linguistic errors, after age 4. In sum, there is little evidence that
parents provide feedback for anything other than the truth value and phonology
of their children’s utterances, consistent with the conclusions of Brown and
Hanlon (1970).
Do parents provide noisy feedback for different types of grammatical constructions?
Only Morgan and Travis (1989) present data that allow construction-wise
comparisons. These data suggest that different types of linguistic errors may
receive opposite types of feedback. For example, since overall Adam’s parents
asked more clarification questions following ill-formed utterances (e.g., past tense
errors) than well-formed utterances, Adam might infer that clarification questions
indicate grammatical errors. But Adam’s parents asked more clarification ques-
tions after well-formed wh- questions (questions containing who, what, where, when, which or why). If Adam eliminated those sentences to which his parents
replied with clarification questions, he could reduce his past tense errors only to
the detriment of his use of wh- questions, or vice versa. To use the parental
feedback in clarification questions, Adam would need to learn different response
contingencies for different grammatical forms; it is far from clear how Adam
could figure this out. Finally, suppose Adam asked a wh- question with a past
tense form (e.g., Who runned?) and his mother replied with a clarification
question. How would Adam determine whether his mother was correcting his wh-
question syntax or correcting his past tense formation?
Summary
There is no evidence that any type of parental feedback is available widely
enough to obviate the need for specific linguistic mechanisms. Only one type of
noisy feedback (recasts, examined in only one study) was available to all the
children studied. No other pattern is provided consistently. No study showed
patterns of complete, partial, or even noisy feedback available for children over
age 4 despite the fact that children continue to make errors after this age. Every
study except Morgan and Travis (1989) collapses morphological and syntactic
errors; several studies even collapse phonological errors with syntactic errors.
Negative evidence 75
Furthermore, because of dubious practices of averaging data from different
children and from children of different ages, some patterns of feedback may
conceal individual differences showing that certain patterns of feedback were not
available to some children and hence not necessary for language acquisition.
Worse, some alleged types of noisy feedback may be completely artifactual and
never provided to any individual child. The existing evidence for noisy feedback
thus fails to withstand the weight of careful empirical scrutiny and provides no
evidence that noisy feedback is necessary for language acquisition.
Definitional artifacts
Many discourse studies classified parental replies in a manner confounded with
the grammaticality of children’s utterances, hence rendering the correlations
between parental replies and children’s grammaticality possibly artifactual. The
coding criteria for parental replies cause two serious problems. First, many types
of parental replies cannot be classified until one looks at the preceding child
utterance, and these replies are sometimes coded differently depending on the
grammaticality of a child’s utterance (Morgan & Travis, 1989; Valian, in press).
For example, some parental utterances might be coded as repetitions following
a grammatical child utterance, but as expansions following an ungrammatical
child utterance. The child says the identical utterance in (5) and (6), but the
utterance is deemed an expansion or recast in the former, and as a repetition in
the latter:
(5) Expansion or recast C: The ball falled down.
P: The ball fell down.
(6) Repetition C: The ball fell down.
P: The ball fell down.
Second, because parents nearly always speak grammatically (99.44% of their
utterances according to Newport, Gleitman, & Gleitman, 1977), certain parental
reply categories, such as repetitions, are biased to occur more often after
grammatical child utterances, while other types of parental coding categories,
such as expansions, are biased to occur more often after ungrammatical child
utterances. Thus some of the contingencies reported in the discourse studies
reflect the consequences of coding categories virtually defined in terms of the
grammaticality of the child’s utterances, rather than being the result of actual
parental sensitivity to the grammaticality of the child’s utterances (Valian, in
press, makes a similar point).
Consider the following analogy. Suppose that I claim that I can control the
76 G. F. Marcus
color of the sky by my choice of which shirt to wear on a given morning. Every
morning, an impartial observer records whether I wear a green shirt or a blue
shirt, and records sky color relative to my shirt color.
After gathering data, we find an interaction between sky color and shirt color:
if 1 wear a blue shirt then the sky may be the same color, but if I wear a green
shirt then the sky is never the same color as my shirt. One might then conclude
that the sky responds to, or is sensitive to, my choice in shirt color. But this
conclusion is false: the sky is not sensitive to whether I wear a blue shirt or a
green shirt.
The problem lies in the coding criteria for the sky’s “replies”. We code sky
color only with respect to shirt color. A blue sky is coded as “same color” if I
wear blue, but the same blue sky is coded as “different color” when I wear green.
The sky is not sensitive to my shirt’s color; the coding categories are. If we instead
simply record the absolute color of the sky, there would be no artifactual
contingency and we would not be misled: the sky is never green no matter what
shirt I wear, and the sky is equally likely to blue no matter what shirt I wear.
Exact repetitions
Repetitions, like sky color, are classified relationally. There is no sentence that is
a repetition independent of a child’s utterance. And just as the “contingency”
between relative sky color and shirt color is confounded with the fact that the sky
is never green, the “contingency” between exact repetitions and child gram-
maticality (i.e., exact repetitions follow more grammatical than ungrammatical
sentences) is also confounded. The contingency between repetitions and well-
formed child speech is inevitable given that parents speak grammatically. When
children speak ungrammatically, there is virtually no chance that their parents will
repeat their utterances verbatim. Conversely, when children speak grammatically,
their parents may repeat their utterance exactly. Demetras et al.‘s exact repeti-
tions, Penner’s repetitions, and Bohannon and Stanowicz’s exact repetitions are
all tainted in this way.” (See also Valian, in press.)
Similar problems beset the other parental reply categories that showed some
(albeit limited) contingency with grammaticality of children’s speech. Recasts,
expansions and clarification questions also can only be coded in relation to
children’s utterances and hence are inherently confounded with children’s gram-
maticality. For instance, the parental reply The ball fell down is coded as a recast
if the child speaks ungrammatically (e.g., The ball falled down) but as a repetition
rather than a recast if the child speaks grammatically (e.g., The ball fell down).
“Hirsh-Pasek et al.‘s category of repetitions differs from exact repetitions because they include
parental responses which contain small changes.
Negative evidence 77
Thus a contingency between recasts and ungrammatical child speech does not
demonstrate that parents are sensitive to their children’s grammar. Whenever
some aspect of a child’s sentence is relevant to the definition of the adult’s reply,
and that aspect is correlated with grammaticality, any correlation between the
grammaticality of the child’s utterance and the type of adult reply will be
confounded.
Possible objections
Even if the observed patterns of feedback are the result of spurious correlations,
the child could still make use of the parental replies. However, parental replies
such as recasts and expansions appear not to reflect any parental sensitivity to
grammaticality. Parents who are insensitive to children’s grammaticality cannot
give children negative evidence (except through explicit metalinguistic statements
of the form Don’t say X); they can only give children positive evidence. Recasts
and expansions may serve as ideal positive evidence, but they do not tell children
what is not in the language.
Another possibility is that whenever a parent says something differently from
the child, the child should assume he or she has made an error. A child following
such a strategy would surely err: most parental utterances are different from their
children’s utterance even when the child speaks grammatically. For instance, a
child might say I want a cookie and the mother might naturally reply No, you’ve
already had three cookies. Discourse is driven by conversation, not implicit
language lessons. Parents may choose a different word or construction to empha-
size something different, or even change the topic entirely. Children who changed
their grammars every time the parent said something different would radically
damage their languages.
Summary
Coding parental replies relative to the child’s previous utterance, as has been
done often in the discourse studies reviewed here, leads to the possibility that the
resulting correlations do not reflect true parental sensitivity but instead reflect
only constant, non-contingent frequencies of certain kinds of child and parental
behavior.
The only evidence for noisy feedback comes from reply categories that are
definitionally flawed. There are some reply types that are not defined relationally,
such as explicit correction, and failures of comprehension (Brown & Hanlon,
1970). But everybody agrees that these reply types are not contingent on
grammaticality.
78 G. F. Marcus
Discussion
There are three serious problems with the position that parents provide negative
evidence to help their children learn language. First, if such feedback does exist,
it is too noisy to be used in practice. Second, it is not available to all children, not
available at all the relevant ages, and probably not available for many types of
errors. Third, reply categories have often been defined relative to the child’s
preceding utterance, and thus observed correlations between parental replies and
children’s utterances may simply be artifacts of the coding scheme. Any one of
these problems is enough to significantly undermine the position that parents
provide negative evidence to their children is significantly undermined.
Positive evidence, in contrast, suffers from none of these problems; nor would
complete feedback if it existed. If a parent says something once, the child can
assume that it is grammatical; the child need not wait for 85 repetitions. Positive
evidence is available to every child regardless of age for all types of representa-
tion. Further, every parental utterance is positive evidence regardless of the child
utterance. If complete feedback (i.e., negative evidence in the sense used by
learnability theorists) existed, it would also meet every test I have applied. If
parents provided some reply type following all and only ungrammatical sentences,
then a single utterance would be a guaranteed test for ungrammaticality, not the
85 times necessary for noisy feedback. Complete feedback would be available by
definition to all children at all ages for all types of errors and could be coded
independently of children’s utterances. But complete feedback apparently does
not exist - nobody has ever claimed that parents correct all and only grammatical
errors.
The fact that there is independent evidence for internal mechanisms provides a
further argument against negative evidence. Consider errors such as Sarah’s Z
maked it with water. As children acquire the English past tense system, they
sometimes apply the regular past rule (add -ed) to irregular stems (e.g., go, make, or sing), thus producing erroneous past tense forms such as goed or maked (see Marcus et al., 1992). If negative evidence is not available, children must stop
producing these forms through some internal (possibly linguistically specific)
mechanism. Marcus et al. (1992) argue that children follow a principle of
inflectional blocking that prevents the application of a regular rule (add -ed to
form the past tense) whenever a child can retrieve an irregular past tense form for
some stem. For example, if a child attempting to mark the past tense of make retrieves made, the regular rule is blocked. If the child fails to retrieve an
irregular past tense form, the rule applies, and the child creates an incorrect form,
such as muked. As children’s retrieval of the correct past tense forms improves
through positive evidence, overregularization errors disappear.
Internal mechanisms, without the aid of negative evidence, could also elimi-
nate errors such as Mark’s Zfilled the sugars up into the bowl, or a similar error, Z
Negative evidence 19
filled the water into the bowl. In these “locative” errors, children use verbs in
inappropriate syntactic constructions. The child has used the “content” argument
(the water) rather than the container “argument” (the bowl) as the object of the
verb fi11.13 In contrast, it is perfectly acceptable to use pour with its “content”
arguments as the object of the verb, as in I poured the water into the glass.
Negative evidence might drive locative errors out of a child’s grammar, but
without negative evidence children would need to eliminate errors through
internal mechanisms. Gropen et al. (1991) argued that locative errors result when
a universal linking rule, object affectedness, is combined with an improper
semantic representation of the verb Jill. The object affectedness rule states that the
direct object of a verb corresponds to an object that is specified as affected in
some particular way in the semantic representation of a verb. Gropen et al.
argued that in the adult grammar, the semantic representation of the verb fill does
not specify how the content is affected (a glass can be filled by pouring but also by
dipping or bailing), hence the content argument cannot appear as the object. In
contrast, the semantics of pour does specify that its content argument is affected
in a specific way, namely it must move downward in a stream, and hence it may
appear as the direct object of the verb. According to this theory, children always
follow the universal linking rule, but if they have an improper semantic repre-
sentation for a verb such as fiZ1, which does specify the manner of motion of the
content argument (e.g., being poured), then they would allow water to be the
object. Gropen et al. predicted that children who thought that fWs semantics
specified that its content argument was affected (e.g., by selecting pictures of
pouring as exemplifying fill) would be the same children who made syntactic
errors such as saying Fill the water into the sink. On their account, children could
learn the correct semantics using only positive evidence (e.g., hearing jZZ used
without pouring as when a cup is dipped in a punch bowl) and crucially, once the
semantics is learned from positive evidence, the linking rule fixes the syntax
automatically, allowing children to unlearn their errors without negative evidence.
If negative evidence really were generally available, it is unlikely that the
empirical evidence would support the existence of these mechanisms, because
they would be solutions to non-problems. But there is evidence for internal
mechanisms. For example, Marcus et al. (1992) showed that children are more
likely to overregularize verbs to which they have less exposure. As exposure to
correct past tense forms increases, overregularization errors decrease, hence
supporting the inflectional blocking hypothesis. Gropen et al. (1991) found that
semantic errors do indeed correlate with syntactic errors; children that make
semantic errors with fill are the same ones that use the wrong syntactic frames for
jifZ, hence supporting the existence of the object affectedness linking rule. These
“The precise details of the syntax and semantics and their relationship are outside of the scope of
this paper.
80 G. F. Marcus
examples suggest that internal linguistic mechanisms are real and that there is no
need for negative evidence.
Finally, it is important to reject the notion that nativist explanations of
language acquisition depend on the lack of negative evidence. Even if perfect
negative evidence were available, innate constraints on the generalizations which
children make would be necessary because many plausible errors simply never
occur. For instance, children never go through a period where they erroneously
form yes-no questions by moving the first is to the front of the sentence.
Although one can turn The man is hungry into Is the man hungry?, children
never, by a false analogy, turn The man who is hungry is ordering dinner into Is the man who hungry is ordering dinner? (e.g., Chomsky, 1965, 1980; Crain &
Nakayama, 1987). More generally, at every stage of language acquisition -
inferring the meaning of a new word or morpheme, creating a morphological or
syntactic rule, or determining the subcategorization frame of a new verb-the
child can make an infinity of logically possible generalizations, regardless of
whether negative evidence exists. The child simply cannot cycle through all logical
possibilities and check to see what his parents say about each one.
Children do learn languages; but they appear to do so without requiring
negative evidence: complete feedback and partial feedback do not exist. Noisy
feedback, the signal that has been claimed to exist, may largely be an averaging
artifact and is at best available in a fraction of the circumstances in which it would
be needed. Many of the observed patterns of noisy feedback are likely to be
averaging artifacts and definitional artifacts, but even if there were no artifacts,
the child would need to say a sentence such as Fill the little sugars up in the bowl more than 85 times to be sufficiently confident that it was an ungrammatical
sentence. There is no existing evidence that noisy feedback can account for the
unlearning of grammatical errors, and given the inherent weakness of noisy
feedback it is also extremely unlikely that any yet-to-be-discovered interactional
patterns would be adequate to account for language acquisition. These considera-
tions suggest that the problem of accounting for children’s avoidance and recovery
from errors in language acquisition is likely to be explained by the nature of their
internal learning mechanisms. The specific mechanisms I have discussed, such as
inflectional blocking and linking rules, are not the only ones that have been
proposed and may not be empirically correct. But any explanation of language
acquisition that depends on parental feedback seems unlikely to succeed, and
quantitative examinations of interaction patterns in parent-child conversation are
unlikely to shed much light on this important issue.
References
Baker, C.L. (1979). Syntactic theory and the projection principle. Linguistic Inquiry, IO, 533-581. Baker, C.L. (1981). Learnability and the English auxilliary system. In C.L. Baker & J.J. McCarthy
(Eds.) The logical problem of language acquisition. (pp. 296-323). Cambridge. MA: MIT Press.
Negative evidence 81
Baker, C.L., & McCarthy, J.J. (Eds.) (1981). The logical problem of language acquisition. Cam- bridge, MA: MIT Press.
Berwick, R.C. (1985). The acquisition of syntactic knowledge. Cambridge, MA: MIT Press.
Bohannon, J.N., MacWhinney, B., & Snow, C. (1990). No negative evidence revisited: Beyond
learnability or who has to prove what to whom. Developmental Psychology, 26, 221-226. Bohannon, J.N., & Stanowicz, L. (1988). The issue of negative evidence: Adult responses to
children’s language errors. Developmental Psychology, 24, 684-689. Bowerman, M. (1983). How do children avoid constructing an overly general grammar in the absence
of feedback about what is not a sentence? Papers and Reports on Child Language Development, 22, 23-25.
Bowerman, M. (1987). Commentary: Mechanisms of language acquisition. In B. MacWhinney (Ed.),
Mechanisms of language acquisition (pp. 443-466). Hillsdale, NJ: Erlbaum. Bowerman, M. (1988). The “no negative evidence” problem: How do children avoid constructing an
overly general grammar? In J. Hawkins (Ed.), Explaining language universals (pp. 73-101).
Oxford: Basil Blackwell. Braine. M.D.S. (1971). On two types of models of the internalization of grammars. In D.I. Slobin
(Ed.), The ontogenesis of grammar (pp. 153-186). New York: Academic Press.
Brown, R. (1973). A first language: The early stages. Cambridge, MA: Harvard University Press.
Brown, R., & Hanlon, C. (1970). Derivational complexity and order of acquisition on child speech. In
J. Hayes (Ed.), Cognition and the developmenf of language (pp. 11-53). New York: Wiley.
Chomsky, N. (1965). Aspects of a theory of syntax. Cambridge, MA: MIT Press.
Chomsky, N. (1980). Rules and representations. New York: Columbia University Press.
Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris.
Chomsky, N., & Lasnik, H. (1977). Filters and control. Linguistic Inquiry, 8, 425-504. Crain, S., & Nakayama, M. (1987). Structure dependence in grammar formation. Language, 63,
522-543. Demetras, M.J., Post, K.N., & Snow, C.E. (1986). Feedback to first language learners: The role of
repetitions and clarification questions. Journal of Child Language, 13, 275-292. Eisenberg, A. (1982). Language acquisition in cultural perspective. Berkeley doctoral dissertation.
Fodor, J.D., & Crain, S. (1987). Simplicity and generality of rules in language acquisition. In B.
MacWhinney (Ed.), Mechanisms of language acquisition (pp. 35-63). Hillsdale, NJ: Erlbaum.
Gold, E.M. (1967). Language identification in the limit. Information and Control, IO, 447-474. Gordon, P. (1990). Learnability and feedback. Developmental Psychology, 26, 217-220. Grimshaw, J. (1981). Form, function, and the language acquisition device. In CL. Baker & J.J.
McCarthy (Eds.), The logical problem of language acquisition (pp. 165-187). Cambridge, MA: MIT Press.
Grimshaw, J., & Pinker, S. (1989). Positive and negative evidence in language acquisition. (Com-
mentary on D. Lightfoot’s “The child’s trigger experience: ‘Degree-O’ Learnability”.) Behavior- al and Brain Sciences, 12, 341.
Gropen, J., Pinker, S., Hollander, M., & Goldberg, R. (1991). Syntax and semantics in the
acquisition of locative verbs. Journal of Child Language, 18, 115-151. Heath, S.B. (1983). Ways with words. Cambridge, UK: Cambridge University Press.
Hirsh-Pasek, K., Treiman, R., & Schneiderman, M. (1984). Brown and Hanlon revisited: Mothers’ sensitivity to ungrammatical forms. Journal of Child Language, 11, 81-88.
Lasnik, H. (1981). Learnability, restrictiveness, and the evaluation metric. In C.L. Baker & J.J.
McCarthy (Eds.), The logical problem of language acquisition (pp. 1-21). Cambridge, MA: MIT Press.
Lasnik, H. (1989). On certain substitutes for negative data. In R. Matthews & W. Demopoulos
(Eds.), Learnability and linguistic theory (pp. 89-105). Dordrecht: Kluwer/Academic Press.
Marcus, G.F., Pinker. S., Ullman, M., Hollander, M., Rosen, T.J., & Xu. F. (1992). Overregulariza-
tion in language acquisition. Monographs of the Society for Research in Child Development, serial no. 228.
Matthews, R.J., & Demopoulous, W. (Eds.) (1989). Learnability and linguistic theory, Dordrecht: KluweriAcademic Press.
Mazurkewich, I., & White, L. (1984). The acquisition of the dative alternation: Unlearning over- generalizations. Cognition, 16, 261-283.
82 G. F. Marcus
McNeill, D. (1966). Developmental psycholinguistics. In F. Smith & G.A. Miller (Eds.). The genesis of language (pp. 15-84). Cambridge, MA: MIT Press.
Moerk, E. (1991). Positive evidence for negative evidence. First Language, 11, 219-251.
Morgan, J.L., Bonamo, K., & Travis, L.L. (1991). Recasts, corrections and negative evidence: Separate and unequal. Submitted for publication.
Morgan, J.L., & Travis, L.L. (1989). Limits on negative information in language input. Journal of Child Language, 16, 531-552.
Newport, E.L., Gleitman, H., & Gleitman, L.R. (1977). Mother, I’d rather do it myself: Some effects and non-effects of maternal speech style. In C.E. Snow & CA. Ferguson (Eds.), Talking ro children: Language inpuf and acquisition (pp. 109-149). Cambridge, UK: Cambridge Universi-
ty Press.
Ochs, E., & Schieffelin, B.B. (1984). Language acquisition and socialization: Three developmental stories and their implications. In R.A. Shweder & R.A. Levine (Eds.), Culture theory: Essays on mind, self and emotion. Cambridge, UK: Cambridge University Press.
Penner, S. (1987). Parental responses to grammatical and ungrammatical child utterances. Child Development, .5X, 376-384.
Pinker. S. (1979). Formal models of language learning. Cognition, 7, 217-283.
Pinker, S. (1984). Language learnability and language development. Cambridge, MA: Harvard
University Press.
Pinker, S. (1989). Learnability and cognifion: The acquisition of argument structure. Cambridge, MA:
MIT Press.
Pye, C. (1986). Quiche Mayan speech to children. Journal of Child Language, 13, 85-100. Roeper, T. (1981). On the deductive model and the acquisition of produtive morphology. In C.L.
Baker & J.J. McCarthy (Eds.), The logical problem of language acquisition. Cambridge, MA:
MIT Press.
Stembcrger, J.P. (1989). The acquisition of morphology: Analysis of a symbolic model of language
acquisition. Unpublished manuscript, Department of Linguistics, University of Minnesota.
Valian, V.V. (in press). Parental replies: Linguistic status and didactic role. Cambridge, MA: MIT Press.
Wexler, K., & Culicover, P. (1980). Formal principles of language acquisition. Cambridge, MA: MIT Press.
Wexler, K.. & Hamburger, H. (1973). On the insufficiency of surface data for the learning of
transformational languages. In J. Hintikka, J. Moravcsik, & P. Suppes (Eds.), Approaches to natural language. Dordrecht: Reidel.
Wexler, K., & Manzini, M.R. (1987). Parameters and learnability in binding theory. In T. Roeper &
E. Williams (Eds.), Parameter setting (pp. 41-76). Dordrecht: Reidel.
Zwicky, A. (1970). A double regularity in the acquisition of English verb morphology. Working Papers in Linguistics 4. Ohio State University.
Appendix: Definitions of parental reply categories
Hirsh- Pasek, Treiman, and Schneiderman (1984)
Strict repetitions: [were coded if the parent repeated] the child’s exact wording
with the possible exceptions that (a) ill-formed aspects were rendered in their
correct grammatical form as in People lives in Florida- People live in Florida. (b) I and you were appropriately interchanged.
Loose repetitions: [were coded if] (d) content words were replaced with proforms
or vice versa . . (e) modifiers were added or deleted . . . (f) a phrase of the
child’s utterance was repeated without the rest of the utterance . . or (g) the
child’s utterance was embedded with a longer utterance.
Negative evidence 83
Demetras, Post, and Snow (1986)
Approvallcorrection: These utterances are explicit responses to the child’s preced-
ing utterance that either approve or give negative feedback to the child. Words
such as yes, no, that’s right, signal explicit feedback.
Repetitions: Four types of repetitions [were] coded: exact, contracted, expanded,
and extended. The use of deictic forms (e.g., substitution of you for me, come for go) are accepted as repetitions.
Exact repetitions: [were coded if the parental reply was] Exact repetition of what
the child said.
Contracted: [were coded if the parental reply was a] Shortening of the child’s
utterance in any way.
Expanded repetitions: [represented a] Correction of child’s utterance with appro- priate syntax or morphology. Child: daddy house/Mother: Daddy’s house.
Clarification questions: These responses refer directly to the child’s preceding
utterance. Questions that start a new topic are not included. Also any
questions requesting knowledge from the child are not included. The different
types of questions are as follows:
Wh- [questions]: . . . must start with a wh- word and clarify, otherwise scored as a
Move-On.
Occasional questions: Questions that have a wh- word embedded in them: You
went where?
Repetition questions: These are repetitions (as described above) that have a rising
intonation contour at the end of the sentence.
Move-ons (MO): In these utterances the mother uses the same topic or starts a
new topic, but does not ‘negotiate’ with the child for meaning. We infer that
the mother understands what the child said, accepts it, and moves on with the
conversation.
Penner (1987)
Topic extensions: . . . continued the current topic of the interaction, but did not
qualify for the placement in another category, such as repetitions or expan-
sions. The topic was defined by the nonverbal as well as the verbal context of
the interaction.
84 G. F. Marcus
Verbal agreementiapproval: Included confirmations, agreements, or praise using
words like “Yes”, “That’s right”, “Good” and “OK”.
Expansions: Parents repeated all or part of the utterance and made additions and
other grammatical or semantic changes to words and morphemes in the
utterance. Parental responses that qualified as expansions were judged to
function to expand upon the previous child’s utterance (e.g., C: “Ball fall” A:
“The ball fell down”).
Repetitions: The parent repeated all or part of the child’s utterance without
adding to the utterance.
No response: The child’s utterance was followed by a pause of at least 2 sec. that
did not contain a verbal or nonverbal parental response or another child’s
utterance.
Confirming questions: included expansions and repetitions that were accompanied
by question intonation. Therefore, this category contained a subset of the
responses that were also defined as either expansions or repetitions.
Bohannon and Stanowicz (1988)
Exact Repetitions: Consisted of verbatim reproductions of the child’s entire
preceding utterance.
Contracted Repetitions: Consisted of the reproduction of a reduced set of ele-
ments from the child’s preceding utterance.
Expanded repetitions were coded if the adult reproduced major elements of the
child’s utterance and added new information.
Recasts [were coded] if the adult preserved the child’s meaning but replaced
elements of the child’s utterance (e.g., C: “That be monkey”; A: “That is a
monkey”).
Clarification questions were . . questions that related to the children’s previous
utterance without requesting any new information and counted other adult
repetitions with rising terminal intonation as exact, contracted, recasted, and
expanded.
Negative evidence 85
Morgan and Travis (1989)
No Response: . . . if one or more child utterances immediately followed the
utterance containing the error and if no utterance in the subsequent adult
conversational turn was explicitly related (via complete or partial imitation) to
the error-containing utterance.
Expansions: If any utterance in the adult conversational turn following the
error. . . expanded the child utterance (e.g., C: Where other stick? P: Where is
the stick?).
Imitations: If the adult exactly repeated the erroneous utterance (reversals in
pronouns and deictic terms notwithstanding).
Clarification questions: If any adult utterance had the force of requesting the child
to repeat all or part of the erroneous utterance.
Confirmation questions: If the adult reply to the erroneous utterance was a
yes-no question pertaining to the linguistic content of the error-containing
utterance.
Move-on: If none of the other categories were applicable.